Deep High-Resolution Representation Learning for Human Pose Estimation
Ke Sun1,2∗† Bin Xiao2∗ Dong Liu1 Jingdong Wang2‡
1
University of Science and Technology of China 2 Microsoft Research Asia
sunk@mail.ustc.edu.cn, dongleiu@ustc.edu.cn, {Bin.Xiao,jingdw}@microsoft.com
Abstract
In this paper, we are interested in the human pose es-
timation problem with a focus on learning reliable high-
resolution representations. Most existing methods recover
high-resolution representations from low-resolution repre-
sentations produced by a high-to-low resolution network.
Instead, our proposed network maintains high-resolution
representations through the whole process.
We start from a high-resolution subnetwork as the first
stage, gradually add high-to-low resolution subnetworks
one by one to form more stages, and connect the mutli-
resolution subnetworks in parallel. We conduct repeated
multi-scale fusions such that each of the high-to-low reso-
lution representations receives information from other par-
allel representations over and over, leading to rich high-
resolution representations. As a result, the predicted key-
point heatmap is potentially more accurate and spatially
more precise. We empirically demonstrate the effectiveness
of our network through the superior pose estimation results
over two benchmark datasets: the COCO keypoint detection
dataset and the MPII Human Pose dataset. In addition, we
show the superiority of our network in pose tracking on the
PoseTrack dataset. The code and models have been publicly
available at https://github.com/leoxiaobin/
deep-high-resolution-net.pytorch.
1. Introduction
2D human pose estimation has been a fundamental yet
challenging problem in computer vision. The goal is to lo-
calize human anatomical keypoints (e.g., elbow, wrist, etc.)
or parts. It has many applications, including human action
recognition, human-computer interaction, animation, etc.
This paper is interested in single-person pose estimation,
which is the basis of other related problems, such as multi-
person pose estimation [6, 26, 32, 38, 46, 55, 40, 45, 17, 69],
∗ Equal
contribution.
† This work is done when Ke Sun was an intern at Microsoft Research,
Beijing, P.R. China
‡ Corresponding author
5693
depth
1×
scale
2×
4×
feature conv. down up
maps unit samp. samp.
Figure 1. Illustrating the architecture of the proposed HRNet. It
consists of parallel high-to-low resolution subnetworks with re-
peated information exchange across multi-resolution subnetworks
(multi-scale fusion). The horizontal and vertical directions cor-
respond to the depth of the network and the scale of the feature
maps, respectively.
video pose estimation and tracking [48, 70], etc.
The recent developments show that deep convolutional
neural networks have achieved the state-of-the-art perfor-
mance. Most existing methods pass the input through a
network, typically consisting of high-to-low resolution sub-
networks that are connected in series, and then raise the
resolution. For instance, Hourglass [39] recovers the high
resolution through a symmetric low-to-high process. Sim-
pleBaseline [70] adopts a few transposed convolution layers
for generating high-resolution representations. In addition,
dilated convolutions are also used to blow up the later lay-
ers of a high-to-low resolution network (e.g., VGGNet or
ResNet) [26, 74].
We present a novel architecture, namely High-
Resolution Net (HRNet), which is able to maintain high-
resolution representations through the whole process. We
start from a high-resolution subnetwork as the first stage,
gradually add high-to-low resolution subnetworks one by
one to form more stages, and connect the multi-resolution
subnetworks in parallel. We conduct repeated multi-scale
fusions by exchanging the information across the paral-
lel multi-resolution subnetworks over and over through the
whole process. We estimate the keypoints over the high-
resolution representations output by our network. The re-
sulting network is illustrated in Figure 1.
(a) (b)

(d)

feature reg. dilated strided up trans. sum

(c) maps conv. conv. conv. samp. conv.

Figure 2. Illustration of representative pose estimation networks that rely on the high-to-low and low-to-high framework. (a) Hourglass [39].
(b) Cascaded pyramid networks [11]. (c) SimpleBaseline [70]: transposed convolutions for low-to-high processing. (d) Combination with
dilated convolutions [26]. Bottom-right legend: reg. = regular convolution, dilated = dilated convolution, trans. = transposed convolution,
strided = strided convolution, concat. = concatenation. In (a), the high-to-low and low-to-high processes are symmetric. In (b), (c) and
(d), the high-to-low process, a part of a classification network (ResNet or VGGNet), is heavy, and the low-to-high process is light. In (a)
and (b), the skip-connections (dashed lines) between the same-resolution layers of the high-to-low and low-to-high processes mainly aim
to fuse low-level and high-level features. In (b), the right part, refinenet, combines the low-level and high-level features that are processed
through convolutions.

Our network has two benefits in comparison to exist-
ing widely-used networks [39, 26, 74, 70] for pose estima-
tion. (i) Our approach connects high-to-low resolution sub-
networks in parallel rather than in series as done in most
existing solutions. Thus, our approach is able to main-
tain the high resolution instead of recovering the resolu-
tion through a low-to-high process, and accordingly the pre-
dicted heatmap is potentially spatially more precise. (ii)
Most existing fusion schemes aggregate low-level and high-
level representations. Instead, we perform repeated multi-
scale fusions to boost the high-resolution representations
with the help of the low-resolution representations of the
same depth and similar level, and vice versa, resulting in
that high-resolution representations are also rich for pose
estimation. Consequently, our predicted heatmap is poten-
tially more accurate.
We empirically demonstrate the superior keypoint detec-
tion performance over two benchmark datasets: the COCO
keypoint detection dataset [35] and the MPII Human Pose
dataset [2]. In addition, we show the superiority of our net-
work in video pose tracking on the PoseTrack dataset [1].
2. Related Work
Most traditional solutions to single-person pose estima-
tion adopt the probabilistic graphical model or the picto-
rial structure model [76, 49], which is recently improved by
exploiting deep learning for better modeling the unary and
pair-wise energies [9, 63, 44] or imitating the iterative in-
ference process [13]. Nowadays, deep convolutional neural
network provides dominant solutions [20, 34, 60, 41, 42,
47, 56, 16]. There are two mainstream methods: regressing
the position of keypoints [64, 7], and estimating keypoint
heatmaps [13, 14, 75] followed by choosing the locations
with the highest heat values as the keypoints.
Most convolutional neural networks for keypoint
heatmap estimation consist of a stem subnetwork similar to
the classification network, which decreases the resolution,
a main body producing the representations with the same
resolution as its input, followed by a regressor estimating
the heatmaps where the keypoint positions are estimated
and then transformed in the full resolution. The main body
mainly adopts the high-to-low and low-to-high framework,
possibly augmented with multi-scale fusion and intermedi-
ate (deep) supervision.
High-to-low and low-to-high. The high-to-low process
aims to generate low-resolution and high-level representa-
tions, and the low-to-high process aims to produce high-
resolution representations [4, 11, 22, 70, 39, 60]. Both the
two processes are possibly repeated several times for boost-
ing the performance [74, 39, 14].
Representative network design patterns include: (i) Sym-
metric high-to-low and low-to-high processes. Hourglass
and its follow-ups [39, 14, 74, 30] design the low-to-high
process as a mirror of the high-to-low process. (ii) Heavy
high-to-low and light low-to-high. The high-to-low pro-
cess is based on the ImageNet classification network, e.g.,

5694
ResNet adopted in [11, 70], and the low-to-high process is
simply a few bilinear-upsampling [11] or transpose convo-
lution [70] layers. (iii) Combination with dilated convolu-
tions. In [26, 50, 34], dilated convolutions are adopted in
the last two stages in the ResNet or VGGNet to eliminate
the spatial resolution loss, which is followed by a light low-
to-high process to further increase the resolution, avoiding
expensive computation cost for only using dilated convolu-
tions [11, 26, 50]. Figure 2 depicts four representative pose
estimation networks.
Multi-scale fusion. The straightforward way is to feed
multi-resolution images separately into multiple networks
and aggregate the output response maps [62]. Hour-
glass [39] and its extensions [74, 30] combine low-level
features in the high-to-low process into the same-resolution
high-level features in the low-to-high process progres-
sively through skip connections. In cascaded pyramid net-
work [11], a globalnet combines low-to-high level features
in the high-to-low process progressively into the low-to-
high process, and then a refinenet combines the low-to-high
level features that are processed through convolutions. Our
approach repeats multi-scale fusion, which is partially in-
spired by deep fusion and its extensions [65, 71, 57, 77, 79].
Intermediate supervision. Intermediate supervision or
deep supervision, early developed for image classifica-
tion [33, 59], is also adopted for helping deep networks
training and improving the heatmap estimation quality,
e.g., [67, 39, 62, 3, 11]. The hourglass approach [39] and
the convolutional pose machine approach [67] process the
intermediate heatmaps as the input or a part of the input of
the remaining subnetwork.
Our approach. Our network connects high-to-low sub-
networks in parallel. It maintains high-resolution repre-
sentations through the whole process for spatially precise
heatmap estimation. It generates reliable high-resolution
representations through repeatedly fusing the representa-
tions produced by the high-to-low subnetworks. Our ap-
proach is different from most existing works, which need
a separate low-to-high upsampling process and aggregate
low-level and high-level representations. Our approach,
without using intermediate heatmap supervision, is superior
in keypoint detection accuracy and efficient in computation
complexity and parameters.
There are related multi-scale networks for classification
and segmentation [5, 8, 72, 78, 29, 73, 53, 54, 23, 80,
53, 51, 18]. Our work is partially inspired by some of
them [54, 23, 80, 53], and there are clear differences making
them not applicable to our problem. Convolutional neural
fabrics [54] and interlinked CNN [80] fail to produce high-
quality segmentation results because of a lack of proper de-
sign on each subnetwork (depth, batch normalization) and
multi-scale fusion. The grid network [18], a combination
5695
of many weight-shared U-Nets, consists of two separate fu-
sion processes across multi-resolution representations: on
the first stage, information is only sent from high resolution
to low resolution; on the second stage, information is only
sent from low resolution to high resolution, and thus less
competitive. Multi-scale densenets [23] does not target and
cannot generate reliable high-resolution representations.
3. Approach
Human pose estimation, a.k.a. keypoint detection, aims
to detect the locations of K keypoints or parts (e.g., elbow,
wrist, etc) from an image I of size W × H × 3. The state-
of-the-art methods transform this problem to estimating K
′ ′
heatmaps of size W ×H , {H1 , H2 , . . . , HK }, where each
heatmap Hk indicates the location confidence of the kth
keypoint.
We follow the widely-adopted pipeline [39, 70, 11] to
predict human keypoints using a convolutional network,
which is composed of a stem consisting of two strided con-
volutions decreasing the resolution, a main body outputting
the feature maps with the same resolution as its input fea-
ture maps, and a regressor estimating the heatmaps where
the keypoint positions are chosen and transformed to the
full resolution. We focus on the design of the main body
and introduce our High-Resolution Net (HRNet) that is de-
picted in Figure 1.
Sequential multi-resolution subnetworks. Existing net-
works for pose estimation are built by connecting high-to-
low resolution subnetworks in series, where each subnet-
work, forming a stage, is composed of a sequence of con-
volutions and there is a down-sample layer across adjacent
subnetworks to halve the resolution.
Let Nsr be the subnetwork in the sth stage and r be the
1
resolution index (Its resolution is 2r−1 of the resolution of
the first subnetwork). The high-to-low network with S (e.g.,
4) stages can be denoted as:
N11 → N22 → N33 → N44 . (1)
Parallel multi-resolution subnetworks. We start from a
high-resolution subnetwork as the first stage, gradually add
high-to-low resolution subnetworks one by one, forming
new stages, and connect the multi-resolution subnetworks
in parallel. As a result, the resolutions for the parallel sub-
networks of a later stage consists of the resolutions from the
previous stage, and an extra lower one.
An example network structure, containing 4 parallel sub-
networks, is given as follows,
N11 → N21 → N31 → N41
ց N22 → N32 → N42
(2)
ց N33 → N43
ց N44 .
 feature
maps
strided
3×3
up samp.
1×1
Figure 3. Illustrating how the exchange unit aggregates the infor-
mation for high, medium and low resolutions from the left to the
right, respectively. Right legend: strided 3 × 3 = strided 3 × 3 con-
volution, up samp. 1×1 = nearest neighbor up-sampling following
a 1 × 1 convolution.
Repeated multi-scale fusion. We introduce exchange units
across parallel subnetworks such that each subnetwork re-
peatedly receives the information from other parallel sub-
networks. Here is an example showing the scheme of ex-
changing information. We divided the third stage into sev-
eral (e.g., 3) exchange blocks, and each block is composed
of 3 parallel convolution units with an exchange unit across
the parallel units, which is given as follows,
1
C31 ց 2
ր C31 ց 3
ր C31 ց last three stages, respectively. The widths of other three
1
C32 → E31 2
→ C32 → E32 3
→ C32 → E33 ,
1
C33 ր 2
ց C33 ր 3
ց C33 ր 96, 192, 384 for HRNet-W48.
(3)
b
where Csr represents the convolution unit in the rth resolu-
tion of the bth block in the sth stage, and Esb is the corre-
sponding exchange unit.
We illustrate the exchange unit in Figure 3 and present
the formulation in the following. We drop the subscript s
and the superscript b for discussion convenience. The in-
puts are s response maps: {X1 , X2 , . . . , Xs }. The outputs
are s response maps: {Y1 , Y2 , . . . , Ys }, whose resolutions
and widths are the same to the input. Each
Ps output is an ag-
gregation of the input maps, Yk = i=1 a(Xi , k). The
exchange unit across stages has an extra output map Ys+1 :
Ys+1 = a(Ys , s + 1).
The function a(Xi , k) consists of upsampling or down-
sampling Xi from resolution i to resolution k. We adopt
strided 3 × 3 convolutions for downsampling. For instance,
one strided 3×3 convolution with the stride 2 for 2× down-
sampling, and two consecutive strided 3 × 3 convolutions
with the stride 2 for 4× downsampling. For upsampling,
we adopt the simple nearest neighbor sampling following a
1 × 1 convolution for aligning the number of channels. If
i = k, a(·, ·) is just an identify connection: a(Xi , k) = Xi .
Heatmap estimation. We regress the heatmaps simply
from the high-resolution representations output by the last
exchange unit, which empirically works well. The loss
function, defined as the mean squared error, is applied
for comparing the predicted heatmaps and the groundtruth
heatmaps. The groundtruth heatmpas are generated by ap-
plying 2D Gaussian with standard deviation of 1 pixel cen-
tered on the grouptruth location of each keypoint.
5696
Network instantiation. We instantiate the network for key-
point heatmap estimation by following the design rule of
ResNet to distribute the depth to each stage and the number
of channels to each resolution.
The main body, i.e., our HRNet, contains four stages
with four parallel subnetworks, whose the resolution is
gradually decreased to a half and accordingly the width (the
number of channels) is increased to the double. The first
stage contains 4 residual units where each unit, the same to
the ResNet-50, is formed by a bottleneck with the width 64,
and is followed by one 3×3 convolution reducing the width
of feature maps to C. The 2nd, 3rd, 4th stages contain 1, 4,
3 exchange blocks, respectively. One exchange block con-
tains 4 residual units where each unit contains two 3 × 3
convolutions in each resolution and an exchange unit across
resolutions. In summary, there are totally 8 exchange units,
i.e., 8 multi-scale fusions are conducted.
In our experiments, we study one small net and one big
net: HRNet-W32 and HRNet-W48, where 32 and 48 rep-
resent the widths (C) of the high-resolution subnetworks in
parallel subnetworks are 64, 128, 256 for HRNet-W32, and
4. Experiments
4.1. COCO Keypoint Detection
Dataset. The COCO dataset [35] contains over 200, 000
images and 250, 000 person instances labeled with 17 key-
points. We train our model on COCO train2017 dataset, in-
cluding 57K images and 150K person instances. We eval-
uate our approach on the val2017 set and test-dev2017 set,
containing 5000 images and 20K images, respectively.
Evaluation metric. The standard evaluation metric is
based
P
on Object Keypoint Similarity (OKS): OKS =
2
i exp(−d /2s2 ki2 )δ(vi >0)
Pi . Here di is the Euclidean dis-
i δ(v i >0)
tance between the detected keypoint and the correspond-
ing ground truth, vi is the visibility flag of the ground
truth, s is the object scale, and ki is a per-keypoint
constant that controls falloff. We report standard aver-
age precision and recall scores1 : AP50 (AP at OKS =
0.50) AP75 , AP (the mean of AP scores at 10 posi-
tions, OKS = 0.50, 0.55, . . . , 0.90, 0.95; APM for medium
objects, APL for large objects, and AR at OKS =
0.50, 0.55, . . . , 0.90, 0.95.
Training. We extend the human detection box in height
or width to a fixed aspect ratio: height : width = 4 : 3,
and then crop the box from the image, which is resized to
a fixed size, 256 × 192 or 384 × 288. The data augmenta-
tion includes random rotation ([−45◦ , 45◦ ]), random scale
1 http://cocodataset.org/#keypoints-eval
Table 1. Comparisons on the COCO validation set. Pretrain = pretrain the backbone on the ImageNet classification task. OHKM = online
hard keypoints mining [11].
Method Backbone Pretrain Input size #Params GFLOPs AP AP50 AP75 APM APL AR
8-stage Hourglass [39] 8-stage Hourglass N 256 × 192 25.1M 14.3 66.9 − − − − −
CPN [11] ResNet-50 Y 256 × 192 27.0M 6.20 68.6 − − − − −
CPN + OHKM [11] ResNet-50 Y 256 × 192 27.0M 6.20 69.4 − − − − −
SimpleBaseline [70] ResNet-50 Y 256 × 192 34.0M 8.90 70.4 88.6 78.3 67.1 77.2 76.3
SimpleBaseline [70] ResNet-101 Y 256 × 192 53.0M 12.4 71.4 89.3 79.3 68.1 78.1 77.1
SimpleBaseline [70] ResNet-152 Y 256 × 192 68.6M 15.7 72.0 89.3 79.8 68.7 78.9 77.8
HRNet-W32 HRNet-W32 N 256 × 192 28.5M 7.10 73.4 89.5 80.7 70.2 80.1 78.9
HRNet-W32 HRNet-W32 Y 256 × 192 28.5M 7.10 74.4 90.5 81.9 70.8 81.0 79.8
HRNet-W48 HRNet-W48 Y 256 × 192 63.6M 14.6 75.1 90.6 82.2 71.5 81.8 80.4
SimpleBaseline [70] ResNet-152 Y 384 × 288 68.6M 35.6 74.3 89.6 81.1 70.5 79.7 79.7
HRNet-W32 HRNet-W32 Y 384 × 288 28.5M 16.0 75.8 90.6 82.7 71.9 82.8 81.0
HRNet-W48 HRNet-W48 Y 384 × 288 63.6M 32.9 76.3 90.8 82.9 72.3 83.4 81.2

Table 2. Comparisons on the COCO test-dev set. #Params and FLOPs are calculated for the pose estimation network, and those for human
detection and keypoint grouping are not included.
Method Backbone Input size #Params GFLOPs AP AP50 AP75 APM APL AR
Bottom-up: keypoint detection and grouping
OpenPose [6] − − − − 61.8 84.9 67.5 57.1 68.2 66.5
Associative Embedding [38] − − − − 65.5 86.8 72.3 60.6 72.6 70.2
PersonLab [45] − − − − 68.7 89.0 75.4 64.1 75.5 75.4
MultiPoseNet [32] − − − − 69.6 86.3 76.6 65.0 76.3 73.5
Top-down: human detection and single-person keypoint detection
Mask-RCNN [21] ResNet-50-FPN − − − 63.1 87.3 68.7 57.8 71.4 −
G-RMI [46] ResNet-101 353 × 257 42.6M 57.0 64.9 85.5 71.3 62.3 70.0 69.7
Integral Pose Regression [58] ResNet-101 256 × 256 45.0M 11.0 67.8 88.2 74.8 63.9 74.0 −
G-RMI + extra data [46] ResNet-101 353 × 257 42.6M 57.0 68.5 87.1 75.5 65.8 73.3 73.3
CPN [11] ResNet-Inception 384 × 288 − − 72.1 91.4 80.0 68.7 77.2 78.5
RMPE [17] PyraNet [74] 320 × 256 28.1M 26.7 72.3 89.2 79.1 68.0 78.6 −
CFN [24] − − − − 72.6 86.1 69.7 78.3 64.1 −
CPN (ensemble) [11] ResNet-Inception 384 × 288 − − 73.0 91.7 80.9 69.5 78.1 79.0
SimpleBaseline [70] ResNet-152 384 × 288 68.6M 35.6 73.7 91.9 81.1 70.3 80.0 79.0
HRNet-W32 HRNet-W32 384 × 288 28.5M 16.0 74.9 92.5 82.8 71.3 80.9 80.1
HRNet-W48 HRNet-W48 384 × 288 63.6M 32.9 75.5 92.5 83.3 71.9 81.5 80.5
HRNet-W48 + extra data HRNet-W48 384 × 288 63.6M 32.9 77.0 92.7 84.5 73.4 83.1 82.0

([0.65, 1.35]), and flipping. Following [66], half body data
augmentation is also involved.
We use the Adam optimizer [31]. The learning sched-
ule follows the setting [70]. The base learning rate is set as
1e−3, and is dropped to 1e−4 and 1e−5 at the 170th and
200th epochs, respectively. The training process is termi-
nated within 210 epochs.
Testing. The two-stage top-down paradigm similar as [46,
11, 70] is used: detect the person instance using a person
detector, and then predict detection keypoints.
We use the same person detectors provided by Simple-
Baseline2 for both validation set and test-dev set. Follow-
2 https://github.com/Microsoft/
human-pose-estimation.pytorch
ing [70, 39, 11], we compute the heatmap by averaging the
headmaps of the original and flipped images. Each keypoint
location is predicted by adjusting the highest heatvalue lo-
cation with a quarter offset in the direction from the highest
response to the second highest response.
Results on the validation set. We report the results of our
method and other state-of–the-art methods in Table 1. Our
small network - HRNet-W32, trained from scratch with the
input size 256 × 192, achieves an 73.4 AP score, outper-
forming other methods with the same input size. (i) Com-
pared to Hourglass [39], our small network improves AP by
6.5 points, and the GFLOPs of our network is much lower
and less than half, while the number of parameters are sim-
ilar and ours is slightly larger. (ii) Compared to CPN [11]

5697
w/o and w/ OHKM, our network, with slightly larger model Table 3. Comparisons on the MPII test set (PCKh@0.5).
size and slightly higher complexity, achieves 4.8 and 4.0 Method Hea Sho Elb Wri Hip Kne Ank Total
points gain, respectively. (iii) Compared to the previous Insafutdinov et al. [26] 96.8 95.2 89.3 84.4 88.4 83.4 78.0 88.5
best-performed SimpleBaseline [70], our HRNet-W32 ob- Wei et al. [67] 97.8 95.0 88.7 84.0 88.4 82.8 79.4 88.5
tains significant improvements: 3.0 points gain for the back- Bulat et al. [4] 97.9 95.1 89.9 85.3 89.4 85.7 81.7 89.7
bone ResNet-50 with a similar model size and GFLOPs, and Newell et al. [39] 98.2 96.3 91.2 87.1 90.1 87.4 83.6 90.9
1.4 points gain for the backbone ResNet-152 whose model Sun et al. [56] 98.1 96.2 91.2 87.2 89.8 87.4 84.1 91.0
size (#Params) and GFLOPs are twice as many as ours. Tang et al. [61] 97.4 96.4 92.1 87.7 90.2 87.7 84.3 91.2
Our nets can benefit from (i) training from the model pre- Ning et al. [43] 98.1 96.3 92.2 87.8 90.6 87.6 82.7 91.2
trained on the ImageNet: The gain is 1.0 points for HRNet- Luvizon et al. [36] 98.1 96.6 92.0 87.5 90.6 88.0 82.7 91.2
W32; (ii) increasing the capacity by increasing the width: Chu et al. [14] 98.5 96.3 91.9 88.1 90.6 88.0 85.0 91.5
HRNet-W48 gets 0.7 and 0.5 points gain for the input sizes Chou et al. [12] 98.2 96.8 92.2 88.0 91.3 89.1 84.9 91.8
256 × 192 and 384 × 288, respectively. Chen et al. [10] 98.1 96.5 92.5 88.5 90.2 89.6 86.0 91.9
Considering the input size 384 × 288, our HRNet-W32 Yang et al. [74] 98.5 96.7 92.5 88.7 91.1 88.6 86.0 92.0
and HRNet-W48, get the 75.8 and 76.3 AP, which have 1.4 Ke et al. [30] 98.5 96.8 92.7 88.4 90.6 89.3 86.3 92.1
and 1.2 improvements compared to the input size 256 × Tang et al. [60] 98.4 96.9 92.6 88.7 91.8 89.4 86.2 92.3
192. In comparison to the SimpleBaseline [70] that uses SimpleBaseline [70] 98.5 96.6 91.9 87.6 91.1 88.1 84.1 91.5
ResNet-152 as the backbone, our HRNet-W32 and HRNet- HRNet-W32 98.6 96.9 92.8 89.0 91.5 89.0 85.7 92.3
W48 attain 1.5 and 2.0 points gain in terms of AP at 45%
Table 4. #Params and GFLOPs of some top-performed methods
and 92.4% computational cost, respectively.
reported in Table 3. The GFLOPs is computed with the input size
Results on the test-dev set. Table 2 reports the pose estima- 256 × 256.
tion performances of our approach and the existing state-of- Method #Params GFLOPs PCKh@0.5
the-art approaches. Our approach is significantly better than Insafutdinov et al. [26] 42.6M 41.2 88.5
bottom-up approaches. On the other hand, our small net- Newell et al. [39] 25.1M 19.1 90.9
work, HRNet-W32, achieves an AP of 74.9. It outperforms Yang et al. [74] 28.1M 21.3 92.0
all the other top-down approaches, and is more efficient in Tang et al. [60] 15.5M 15.6 92.3
terms of model size (#Params) and computation complexity SimpleBaseline [70] 68.6M 20.9 91.5
(GFLOPs). Our big model, HRNet-W48, achieves the high- HRNet-W32 28.5M 9.5 92.3
est 75.5 AP. Compared to the SimpleBaseline [70] with the
same input size, our small and big networks receive 1.2 and ground-truth head bounding box. The PCKh@0.5 (α = 0.5)
1.8 improvements, respectively. With additional data from score is reported.
AI Challenger [68] for training, our single big network can
Results on the test set. Tables 3 and 4 show the
obtain an AP of 77.0.
PCKh@0.5 results, the model size and the GFLOPs of the
4.2. MPII Human Pose Estimation top-performed methods. We reimplement the SimpleBase-
line [70] by using ResNet-152 as the backbone with the
Dataset. The MPII Human Pose dataset [2] consists of im- input size 256 × 256. Our HRNet-W32 achieves a 92.3
ages taken from a wide-range of real-world activities with PCKh@0.5 score, and outperforms the stacked hourglass
full-body pose annotations. There are around 25K images approach [39] and its extensions [56, 14, 74, 30, 60]. Our
with 40K subjects, where there are 12K subjects for test- result is the same as the best one [60] among the previously-
ing and the remaining subjects for the training set. The data published results on the leaderboard of Nov. 16th, 20183 .
augmentation and the training strategy are the same to MS We would like to point out that the approach [60], comple-
COCO, except that the input size is cropped to 256 × 256 mentary to our approach, exploits the compositional model
for fair comparison with other methods. to learn the configuration of human bodies and adopts multi-
Testing. The testing procedure is almost the same to that level intermediate supervision, from which our approach
in COCO except that we adopt the standard testing strategy can also benefit. We also tested our big network - HRNet-
to use the provided person boxes instead of detected person W48 and obtained the same result 92.3. The reason might
boxes. Following [14, 74, 60], a six-scale pyramid testing be that the performance in this dataset tends to be saturate.
procedure is performed.
Evaluation metric. The standard metric [2], the PCKh 4.3. Application to Pose Tracking
(head-normalized probability of correct keypoint) score, is Dataset. PoseTrack [27] is a large-scale benchmark for hu-
used. A joint is correct if it falls within αl pixels of the man pose estimation and articulated tracking in video. The
groundtruth position, where α is a constant and l is the head
size that corresponds to 60% of the diagonal length of the 3 http://human-pose.mpi-inf.mpg.de/#results

5698
dataset, based on the raw videos provided by the popular Table 5. Results of pose tracking on the PoseTrack2017 test set.
MPII Human Pose dataset, contains 550 video sequences Entry Additional training Data mAP MOTA
with 66, 374 frames. The video sequences are split into ML-LAB [81] COCO+MPII-Pose 70.3 41.8
292, 50, 208 videos for training, validation, and testing, re- SOPT-PT [52] COCO+MPII-Pose 58.2 42.0
spectively. The length of the training videos ranges between BUTD2 [28] COCO 59.2 50.6
41 − 151 frames, and 30 frames from the center of the video MVIG [52] COCO+MPII-Pose 63.2 50.7
are densely annotated. The number of frames in the valida- PoseFlow [52] COCO+MPII-Pose 63.0 51.0
tion/testing videos ranges between 65 − 298 frames. The ProTracker [19] COCO 59.6 51.8
30 frames around the keyframe from the MPII Pose dataset HMPT [52] COCO+MPII-Pose 63.7 51.9
are densely annotated, and afterwards every fourth frame is JointFlow [15] COCO 63.6 53.1
annotated. In total, this constitutes roughly 23, 000 labeled STAF [52] COCO+MPII-Pose 70.3 53.8
frames and 153, 615 pose annotations. MIPAL [52] COCO 68.8 54.5
Evaluation metric. We evaluate the results from two as- FlowTrack [70] COCO 74.6 57.8
pects: frame-wise multi-person pose estimation, and multi- HRNet-W48 COCO 74.9 57.9
person pose tracking. Pose estimation is evaluated by the
mean Average Precision (mAP) as done in [50, 27]. Multi- Table 6. Ablation study of exchange units that are used in re-
person pose tracking is evaluated by the multi-object track- peated multi-scale fusion. Int. exchange across = intermediate
ing accuracy (MOTA) [37, 27]. Details are given in [27]. exchange across stages, Int. exchange within = intermediate ex-
change within stages.
Training. We train our HRNet-W48 for single person pose Method Final exchange Int. exchange across Int. exchange within AP
estimation on the PoseTrack2017 training set, where the (a) X 70.8
network is initialized by the model pre-trained on COCO (b) X X 71.9
dataset. We extract the person box, as the input of our net- (c) X X X 73.4
work, from the annotated keypoints in the training frames
by extending the bounding box of all the keypoints (for one
datasets. This further implies the effectiveness of our pose
single person) by 15% in length. The training setup, includ-
estimation network.
ing data augmentation, is almost the same as that for COCO
except that the learning schedule is different (as now it is 4.4. Ablation Study
for fine-tuning): the learning rate starts from 1e−4, drops
to 1e−5 at the 10th epoch, and to 1e−6 at the 15th epoch; We study the effect of each component in our approach
the iteration ends within 20 epochs. on the COCO keypoint detection dataset. All results are
obtained over the input size of 256 × 192 except the study
Testing. We follow [70] to track poses across frames. It about the effect of the input size.
consists of three steps: person box detection and propa-
gation, human pose estimation, and pose association cross Repeated multi-scale fusion. We empirically analyze the
nearby frames. We use the same person box detector as used effect of the repeated multi-scale fusion. We study three
in SimpleBaseline [70], and propagate the detected box into variants of our network. (a) W/o intermediate exchange
nearby frames by propagating the predicted keypoints ac- units (1 fusion): There is no exchange between multi-
cording to the optical flows computed by FlowNet 2.0 [25]4 , resolution subnetworks except the last exchange unit. (b)
followed by non-maximum suppression for box removing. W/ across-stage exchange units (3 fusions): There is no ex-
The pose association scheme is based on the object key- change between parallel subnetworks within each stage. (c)
point similarity between the keypoints in one frame and the W/ both across-stage and within-stage exchange units (to-
keypoints propagated from the nearby frame according to tally 8 fusion): This is our proposed method. All the net-
the optical flows. The greedy matching algorithm is then works are trained from scratch. The results on the COCO
used to compute the correspondence between keypoints in validation set given in Table 6 show that the multi-scale fu-
nearby frames. More details are given in [70]. sion is helpful and more fusions lead to better performance.
Results on the PoseTrack2017 test set. Table 5 reports Resolution maintenance. We study the performance of a
the results. Compared with the second best approach, the variant of the HRNet: all the four high-to-low resolution
FlowTrack in SimpleBaseline [70], that uses ResNet-152 subnetworks are added at the beginning and the depth are
as the backbone, our approach gets 0.3 and 0.1 points gain the same; the fusion schemes are the same to ours. Both
in terms of mAP and MOTA, respectively. The superi- our HRNet-W32 and the variant (with similar #Params and
ority over the FlowTrack [70] is consistent to that on the GFLOPs) are trained from scratch and tested on the COCO
COCO keypoint detection and MPII human pose estimation validation set. The variant achieves an AP of 72.5, which
is lower than the 73.4 AP of HRNet-W32. We believe that
4 https://github.com/NVIDIA/flownet2-pytorch the reason is that the low-level features extracted from the

5699
Figure 4. Qualitative results of some example images in the MPII (top) and COCO (bottom) datasets: containing viewpoint and appearance
change, occlusion, multiple persons, and common imaging artifacts.


1x 2x 4x
[

   [
[ [

 
[
[ [
  [
[

$3
$3
 



+51HW: +51HW:
Figure 5. Ablation study of high- and low-resolution representa-
tions. 1×, 2×, 4× correspond to the representations of the high,
medium, low resolutions, respectively.
early stages over the low-resolution subnetworks are less
helpful. In addition, the simple high-resolution network of
similar parameter and GFLOPs without low-resolution par-
allel subnetworks shows much lower performance .
Representation resolution. We study how the represen-
tation resolution affects the pose estimation performance
from two aspects: check the quality of the heatmap esti-
mated from the feature maps of each resolution from high
to low, and study how the input size affects the quality.
We train our small and big networks initialized by the
model pretrained for the ImageNet classification. Our net-
work outputs four response maps from high-to-low solu-
tions. The quality of heatmap prediction over the lowest-
resolution response map is too low and the AP score is be-
low 10 points. The AP scores over the other three maps
are reported in Figure 5. The comparison implies that the
resolution does impact the keypoint prediction quality.
Figure 6 shows how the input size affects the perfor-
mance in comparison with SimpleBaseline (ResNet-50).
Thanks to maintaining the high resolution through the
whole process, the improvement for the smaller input size
is more significant than the larger input size, e.g., the im-
provement is 4.0 points for 256 × 192 and 6.3 points for


 [
[


6LPSOH%DVHOLQH

+51HW:
[

       
*)/23V
Figure 6. Illustrating how the performances of our HRNet and
SimpleBaseline [70] are affected by the input size.
128 × 96. This implies that our approach is more advanta-
geous in the real applications where the computation cost is
also an important factor. On the other hand, our approach
with the input size 256 × 192 outperforms the SimpleBase-
line with the large input size of 384 × 288.
5. Conclusion and Future Works
In this paper, we present a high-resolution network for
human pose estimation, yielding accurate and spatially-
precise keypoint heatmaps. The success stems from two
aspects: (i) maintain the high resolution through the whole
process without the need of recovering the high resolu-
tion; and (ii) fuse multi-resolution representations repeat-
edly, rendering reliable high-resolution representations.
The future works include the applications to other dense
prediction tasks5 , e.g., face alignment, object detection, se-
mantic segmentation, as well as the investigation on aggre-
gating multi-resolution representations in a less light way.
Acknowledgements. The authors thank Dianqi Li and Lei
Zhang for helpful discussions. Dr. Liu was partially sup-
ported by the Strategic Priority Research Program of the
Chinese Academy of Sciences under Grant XDB06040900.
5 The project page is https://jingdongwang2017.github.
io/Projects/HRNet/
5700
References
[1] Mykhaylo Andriluka, Umar Iqbal, Anton Milan, Eldar In-
safutdinov, Leonid Pishchulin, Juergen Gall, and Bernt
Schiele. Posetrack: A benchmark for human pose estima-
tion and tracking. In CVPR, pages 5167–5176, 2018. 2
[2] Mykhaylo Andriluka, Leonid Pishchulin, Peter V. Gehler,
and Bernt Schiele. 2d human pose estimation: New bench-
mark and state of the art analysis. In CVPR, pages 3686–
3693, 2014. 2, 6
[3] Vasileios Belagiannis and Andrew Zisserman. Recurrent hu-
man pose estimgation. In FG, pages 468–475, 2017. 3
[4] Adrian Bulat and Georgios Tzimiropoulos. Human pose esti-
mation via convolutional part heatmap regression. In ECCV,
volume 9911 of Lecture Notes in Computer Science, pages
717–732. Springer, 2016. 2, 6
[5] Zhaowei Cai, Quanfu Fan, Rogério Schmidt Feris, and Nuno
Vasconcelos. A unified multi-scale deep convolutional neu-
ral network for fast object detection. In ECCV, pages 354–
370, 2016. 3
[6] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh.
Realtime multi-person 2d pose estimation using part affinity
fields. In CVPR, pages 1302–1310, 2017. 1, 5
[7] João Carreira, Pulkit Agrawal, Katerina Fragkiadaki, and Ji-
tendra Malik. Human pose estimation with iterative error
feedback. In CVPR, pages 4733–4742, 2016. 2
[8] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos,
Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic im-
age segmentation with deep convolutional nets, atrous con-
volution, and fully connected crfs. IEEE Trans. Pattern Anal.
Mach. Intell., 40(4):834–848, 2018. 3
[9] Xianjie Chen and Alan L. Yuille. Articulated pose estima-
tion by a graphical model with image dependent pairwise
relations. In NIPS, pages 1736–1744, 2014. 2
[10] Yu Chen, Chunhua Shen, Xiu-Shen Wei, Lingqiao Liu, and
Jian Yang. Adversarial posenet: A structure-aware convolu-
tional network for human pose estimation. In ICCV, pages
1221–1230, 2017. 6
[11] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang
Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network
for multi-person pose estimation. CoRR, abs/1711.07319,
2017. 2, 3, 5
[12] Chia-Jung Chou, Jui-Ting Chien, and Hwann-Tzong Chen.
Self adversarial training for human pose estimation. CoRR,
abs/1707.02439, 2017. 6
[13] Xiao Chu, Wanli Ouyang, Hongsheng Li, and Xiaogang
Wang. Structured feature learning for pose estimation. In
CVPR, pages 4715–4723, 2016. 2
[14] Xiao Chu, Wei Yang, Wanli Ouyang, Cheng Ma, Alan L.
Yuille, and Xiaogang Wang. Multi-context attention for hu-
man pose estimation. In CVPR, pages 5669–5678, 2017. 2,
6 [32] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Mul-
[15] Andreas Doering, Umar Iqbal, and Juergen Gall. Joint flow:
Temporal flow fields for multi person tracking, 2018. 7
[16] Xiaochuan Fan, Kang Zheng, Yuewei Lin, and Song Wang.
Combining local appearance and holistic view: Dual-source
deep neural networks for human pose estimation. In CVPR,
pages 1347–1355, 2015. 2
5701
[17] Haoshu Fang, Shuqin Xie, Yu-Wing Tai, and Cewu Lu.
RMPE: regional multi-person pose estimation. In ICCV,
pages 2353–2362, 2017. 1, 5
[18] Damien Fourure, Rémi Emonet, Élisa Fromont, Damien
Muselet, Alain Trémeau, and Christian Wolf. Residual conv-
deconv grid network for semantic segmentation. In British
Machine Vision Conference 2017, BMVC 2017, London, UK,
September 4-7, 2017, 2017. 3
[19] Rohit Girdhar, Georgia Gkioxari, Lorenzo Torresani,
Manohar Paluri, and Du Tran. Detect-and-track: Efficient
pose estimation in videos. In CVPR, pages 350–359, 2018.
7
[20] Georgia Gkioxari, Alexander Toshev, and Navdeep Jaitly.
Chained predictions using convolutional neural networks. In
ECCV, pages 728–743, 2016. 2
[21] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross B.
Girshick. Mask R-CNN. In ICCV, pages 2980–2988, 2017.
5
[22] Peiyun Hu and Deva Ramanan. Bottom-up and top-down
reasoning with hierarchical rectified gaussians. In CVPR,
pages 5600–5609, 2016. 2
[23] Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens
van der Maaten, and Kilian Q. Weinberger. Multi-scale
dense convolutional networks for efficient prediction. CoRR,
abs/1703.09844, 2017. 3
[24] Shaoli Huang, Mingming Gong, and Dacheng Tao. A coarse-
fine network for keypoint localization. In ICCV, pages 3047–
3056. IEEE Computer Society, 2017. 5
[25] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper,
Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolu-
tion of optical flow estimation with deep networks. In CVPR,
pages 1647–1655, 2017. 7
[26] Eldar Insafutdinov, Leonid Pishchulin, Bjoern Andres,
Mykhaylo Andriluka, and Bernt Schiele. Deepercut: A
deeper, stronger, and faster multi-person pose estimation
model. In ECCV, pages 34–50, 2016. 1, 2, 3, 6
[27] Umar Iqbal, Anton Milan, and Juergen Gall. Posetrack: Joint
multi-person pose estimation and tracking. In CVPR, pages
4654–4663, 2017. 6, 7
[28] Sheng Jin, Xujie Ma, Zhipeng Han, Yue Wu, Wei Yang,
Wentao Liu, Chen Qian, and Wanli Ouyang. Towards multi-
person pose tracking: Bottom-up and top-down methods. In
ICCV PoseTrack Workshop, 2017. 7
[29] Angjoo Kanazawa, Abhishek Sharma, and David W. Ja-
cobs. Locally scale-invariant convolutional neural networks.
CoRR, abs/1412.5104, 2014. 3
[30] Lipeng Ke, Ming-Ching Chang, Honggang Qi, and Siwei
Lyu. Multi-scale structure-aware network for human pose
estimation. CoRR, abs/1803.09894, 2018. 2, 3, 6
[31] Diederik P. Kingma and Jimmy Ba. Adam: A method for
stochastic optimization. CoRR, abs/1412.6980, 2014. 5
tiposenet: Fast multi-person pose estimation using pose
residual network. In ECCV, volume 11215 of Lecture Notes
in Computer Science, pages 437–453. Springer, 2018. 1, 5
[33] Chen-Yu Lee, Saining Xie, Patrick W. Gallagher, Zhengyou
Zhang, and Zhuowen Tu. Deeply-supervised nets. In AIS-
TATS, 2015. 3
[34] Ita Lifshitz, Ethan Fetaya, and Shimon Ullman. Human pose
estimation using deep consensus voting. In ECCV, pages
246–260, 2016. 2, 3
[35] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James
Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and
C. Lawrence Zitnick. Microsoft COCO: common objects in
context. In ECCV, pages 740–755, 2014. 2, 4
[36] Diogo C. Luvizon, Hedi Tabia, and David Picard. Human
pose regression by combining indirect part detection and
contextual information. CoRR, abs/1710.02322, 2017. 6
[37] Anton Milan, Laura Leal-Taixé, Ian D. Reid, Stefan Roth,
and Konrad Schindler. MOT16: A benchmark for multi-
object tracking. CoRR, abs/1603.00831, 2016. 7
[38] Alejandro Newell, Zhiao Huang, and Jia Deng. Associa-
tive embedding: End-to-end learning for joint detection and
grouping. In NIPS, pages 2274–2284, 2017. 1, 5
[39] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hour-
glass networks for human pose estimation. In ECCV, pages
483–499, 2016. 1, 2, 3, 5, 6
[40] Xuecheng Nie, Jiashi Feng, Junliang Xing, and Shuicheng
Yan. Pose partition networks for multi-person pose estima-
tion. In ECCV, September 2018. 1
[41] Xuecheng Nie, Jiashi Feng, and Shuicheng Yan. Mutual
learning to adapt for joint human parsing and pose estima-
tion. In ECCV, September. 2
[42] Xuecheng Nie, Jiashi Feng, Yiming Zuo, and Shuicheng
Yan. Human pose estimation with parsing induced learner.
In CVPR, June 2018. 2
[43] Guanghan Ning, Zhi Zhang, and Zhiquan He. Knowledge-
guided deep fractal neural networks for human pose estima-
tion. IEEE Trans. Multimedia, 20(5):1246–1259, 2018. 6
[44] Wanli Ouyang, Xiao Chu, and Xiaogang Wang. Multi-source
deep learning for human pose estimation. In CVPR, pages
2337–2344, 2014. 2
[45] George Papandreou, Tyler Zhu, Liang-Chieh Chen, Spyros
Gidaris, Jonathan Tompson, and Kevin Murphy. Person-
lab: Person pose estimation and instance segmentation with
a bottom-up, part-based, geometric embedding model. In
ECCV, September 2018. 1, 5
[46] George Papandreou, Tyler Zhu, Nori Kanazawa, Alexander
Toshev, Jonathan Tompson, Chris Bregler, and Kevin Mur-
phy. Towards accurate multi-person pose estimation in the
wild. In CVPR, pages 3711–3719, 2017. 1, 5
[47] Xi Peng, Zhiqiang Tang, Fei Yang, Rogerio S. Feris, and
Dimitris Metaxas. Jointly optimize data augmentation and
network training: Adversarial data augmentation in human
pose estimation. In CVPR, June 2018. 2
[48] Tomas Pfister, James Charles, and Andrew Zisserman. Flow-
ing convnets for human pose estimation in videos. In ICCV,
pages 1913–1921, 2015. 1
[49] Leonid Pishchulin, Mykhaylo Andriluka, Peter V. Gehler,
and Bernt Schiele. Poselet conditioned pictorial structures.
In CVPR, pages 588–595, 2013. 2
[50] Leonid Pishchulin, Eldar Insafutdinov, Siyu Tang, Bjoern
Andres, Mykhaylo Andriluka, Peter V. Gehler, and Bernt
Schiele. Deepcut: Joint subset partition and labeling for
multi person pose estimation. In CVPR, pages 4929–4937,
2016. 3, 7
5702
[51] Tobias Pohlen, Alexander Hermans, Markus Mathias, and
Bastian Leibe. Full-resolution residual networks for seman-
tic segmentation in street scenes. In CVPR, 2017. 3
[52] PoseTrack. PoseTrack Leader Board. https://
posetrack.net/leaderboard.php. 7
[53] Mohamed Samy, Karim Amer, Kareem Eissa, Mahmoud
Shaker, and Mohamed ElHelw. Nu-net: Deep residual wide
field of view convolutional neural network for semantic seg-
mentation. In CVPRW, June 2018. 3
[54] Shreyas Saxena and Jakob Verbeek. Convolutional neural
fabrics. In NIPS, pages 4053–4061, 2016. 3
[55] Taiki Sekii. Pose proposal networks. In ECCV, September
2018. 1
[56] Ke Sun, Cuiling Lan, Junliang Xing, Wenjun Zeng, Dong
Liu, and Jingdong Wang. Human pose estimation using
global and local normalization. In ICCV, pages 5600–5608,
2017. 2, 6
[57] Ke Sun, Mingjie Li, Dong Liu, and Jingdong Wang. IGCV3:
interleaved low-rank group convolutions for efficient deep
neural networks. In BMVC, page 101. BMVA Press, 2018. 3
[58] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen
Wei. Integral human pose regression. In ECCV, pages 536–
553, 2018. 5
[59] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet,
Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent
Vanhoucke, and Andrew Rabinovich. Going deeper with
convolutions. In CVPR, pages 1–9, 2015. 3
[60] Wei Tang, Pei Yu, and Ying Wu. Deeply learned composi-
tional models for human pose estimation. In ECCV, Septem-
ber 2018. 2, 6
[61] Zhiqiang Tang, Xi Peng, Shijie Geng, Lingfei Wu, Shaoting
Zhang, and Dimitris N. Metaxas. Quantized densely con-
nected u-nets for efficient landmark localization. In ECCV,
pages 348–364, 2018. 6
[62] Jonathan Tompson, Ross Goroshin, Arjun Jain, Yann LeCun,
and Christoph Bregler. Efficient object localization using
convolutional networks. In CVPR, pages 648–656, 2015. 3
[63] Jonathan J. Tompson, Arjun Jain, Yann LeCun, and
Christoph Bregler. Joint training of a convolutional network
and a graphical model for human pose estimation. In NIPS,
pages 1799–1807, 2014. 2
[64] Alexander Toshev and Christian Szegedy. Deeppose: Human
pose estimation via deep neural networks. In CVPR, pages
1653–1660, 2014. 2
[65] Jingdong Wang, Zhen Wei, Ting Zhang, and Wenjun Zeng.
Deeply-fused nets. CoRR, abs/1605.07716, 2016. 3
[66] Zhicheng Wang, Wenbo Li, Binyi Yin, Qixiang Peng, Tianzi
Xiao, Yuming Du, Zeming Li, Xiangyu Zhang, Gang Yu,
and Jian Sun. Mscoco keypoints challenge 2018. In Joint
Recognition Challenge Workshop at ECCV 2018, 2018. 5
[67] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser
Sheikh. Convolutional pose machines. In CVPR, pages
4724–4732, 2016. 3, 6
[68] Jiahong Wu, He Zheng, Bo Zhao, Yixin Li, Baoming
Yan, Rui Liang, Wenjia Wang, Shipei Zhou, Guosen Lin,
Yanwei Fu, et al. Ai challenger: A large-scale dataset
for going deeper in image understanding. arXiv preprint
arXiv:1711.06475, 2017. 6
[69] Fangting Xia, Peng Wang, Xianjie Chen, and Alan L. Yuille.
Joint multi-person pose estimation and semantic part seg-
mentation. In CVPR, pages 6080–6089, 2017. 1
[70] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines
for human pose estimation and tracking. In ECCV, pages
472–487, 2018. 1, 2, 3, 5, 6, 7, 8
[71] Guotian Xie, Jingdong Wang, Ting Zhang, Jianhuang Lai,
Richang Hong, and Guo-Jun Qi. Interleaved structured
sparse convolutional neural networks. In CVPR, pages 8847–
8856. IEEE Computer Society, 2018. 3
[72] Saining Xie and Zhuowen Tu. Holistically-nested edge de-
tection. In ICCV, pages 1395–1403, 2015. 3
[73] Yichong Xu, Tianjun Xiao, Jiaxing Zhang, Kuiyuan Yang,
and Zheng Zhang. Scale-invariant convolutional neural net-
works. CoRR, abs/1411.6369, 2014. 3
[74] Wei Yang, Shuang Li, Wanli Ouyang, Hongsheng Li, and
Xiaogang Wang. Learning feature pyramids for human pose
estimation. In ICCV, pages 1290–1299, 2017. 1, 2, 3, 5, 6
[75] Wei Yang, Wanli Ouyang, Hongsheng Li, and Xiaogang
Wang. End-to-end learning of deformable mixture of parts
and deep convolutional neural networks for human pose es-
timation. In CVPR, pages 3073–3082, 2016. 2
[76] Yi Yang and Deva Ramanan. Articulated pose estimation
with flexible mixtures-of-parts. In CVPR, pages 1385–1392,
2011. 2
[77] Ting Zhang, Guo-Jun Qi, Bin Xiao, and Jingdong Wang. In-
terleaved group convolutions. In ICCV, pages 4383–4392,
2017. 3
[78] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang
Wang, and Jiaya Jia. Pyramid scene parsing network. In
CVPR, pages 6230–6239, 2017. 3
[79] Liming Zhao, Mingjie Li, Depu Meng, Xi Li, Zhaoxi-
ang Zhang, Yueting Zhuang, Zhuowen Tu, and Jingdong
Wang. Deep convolutional neural networks with merge-and-
run mappings. In IJCAI, pages 3170–3176, 2018. 3
[80] Yisu Zhou, Xiaolin Hu, and Bo Zhang. Interlinked convo-
lutional neural networks for face parsing. In ISNN, pages
222–231, 2015. 3
[81] Xiangyu Zhu, Yingying Jiang, and Zhenbo Luo. Multi-
person pose estimation for posetrack with enhanced part
affinity fields. In ICCV PoseTrack Workshop, 2017. 7

5703